融合关系网络的Tracktor++多目标跟踪算法  

Combining Relation Network in Tracktor++for Multi-target Tracking

作　　者：姬张建 薛冰心 JI Zhangjian;XUE Bingxin(School of Computer and Information Technology,Shanxi University,Taiyuan 030006,China;Institute of Big Data Science and Industry,Shanxi University,Taiyuan 030006,China)

机构地区：[1]山西大学计算机与信息技术学院,山西太原030006 [2]山西大学大数据科学与产业研究院,山西太原030006

出　　处：《山西大学学报（自然科学版）》2023年第5期1076-1084,共9页Journal of Shanxi University(Natural Science Edition)

基　　金：国家自然科学基金(61602288,61703252,61702314);山西省自然科学基金(20210302123443,201901D211176,201901D211170)。

摘　　要：为了提升Tracktor++在多目标跟踪中的性能,提出了一种改进的多目标跟踪方法,即关系网Tracktor++(RN-Tracktor++),它在Tracktor++中引入关系网络进行目标关联。对于给定的视频序列,首先通过FRCNN-FPN(Faster Region Convolutional Neural Network-Feature Pyramid Network)检测器的回归分支从前一帧中的目标框预测得到它们在当前帧的位置,然后通过关系网络将它们与已有轨迹进行关联,从而得到每个特定目标的运动轨迹。在MOT17、MOT20及降采样的PETS(Performance Evaluation of Tracking and Surveillance)数据集上的实验结果表明,所提出的方法实现了更好的跟踪性能,特别是在降采样的PETS数据集上,与Tracktor++相比,MOTA(Multiple Object Tracking Accuracy)精度提高了1.6%,证明了关系网络用于目标关联可提高跟踪性能。In order to improve the performance of Tracktor++for multi-target tracking,an improved multi-target tracking method is proposed,namely RN-Tracktor++,which introduces the relation network into the Tracktor++for target association.For a given vid-eo sequence,first the regression branch of FRCNN-FPN(Faster Region Convolutional Neural Network-Feature Pyramid Network)detector is adopted to predict their new ones in the current frame from the bounding boxes of the targets in the previous frame,and then they are associated with the existing trajectories through the relation network to obtain the motion trajectory of each specific tar-get.The experimental results on three datasets,MOT17,MOT20 and downsampled PETS(Performance Evaluation of Tracking and Surveillance),demonstrate that the proposed method achieves better performance,especially on the downsampled PETS dataset.Compared with Tracktor++,its MOTA(Multiple Object Tracking Accuracy)raises by 1.6%,which verifies that the relation network as a target association method can enhance the tracking performance.

关 键 词：多目标跟踪 目标检测 目标关联 Tracktor++ 关系网络 

分 类 号：O436[机械工程—光学工程]